Kiril Alexiev

Implementation of Hough Transform as track detector

Hough Transform is a convenient tool for features extraction from images. In this paper an implementation of Hough Transform is considered for automatic track initiation in the surveillance radar space. The need of track initiation arises when there are many moving objects in the sensor surveillance volume and when it is not clear which measurement to which target belongs. If the type of target trajectories is known, the trajectories can be easily detected by performing corresponding Hough transform on the track image. Here, the effectiveness of the Hough transform track initiator is discussed. The influence of Hough parameter space granularity upon probability of track detection is analyzed. Analytical expressions for probability of track detection using Hough Transform are derived in the presence of normal distributed additive system noise, measurement noise and without any noises. A new parameter space structure, matching with measurement errors is proposed. The Monte Carlo simulation confirms received analytical result.

Sound fields clusterization via neural networks

Paper presents application of a recently proposed approach for multidimensional data clustering to data received from a microphone array antenna. The accumulated sound pressure at each point (a microphone in the array) is used to create “sound picture” of the observed by the microphone antenna area. Features for classification are extracted using overlapping receptive fields based on the model of direction selective cells in the middle temporal (MT) cortex. Next the clustering procedure using Echo state network and subtractive clustering algorithm is applied to separate receptive fields in proper number of classes. The obtained results are compared with the sonograms created by the original software of the producer of microphone array.

Echo State Networks in Dynamic Data Clustering

The present paper follows the initial work on multidimensional static data clustering using a novel neural network structure, namely Echo state network (ESN). Here we exploit dynamic nature of these networks for solving of clustering task of a multidimensional dynamic data set. The data used in this investigation are taken from an experimental set-up applied for tasting of visual discrimination of complex dot motions. The proposed here model, although far from complicated brain theories, can serve as a good basis for investigation of humans perception.

Recurrent neural networks for automatic clustering of multispectral satellite images

In the present work we applied a recently developed procedure for multidimensional data clustering to multispectral satellite images. The core of our approach lays in projection of the multidimensional image to a two dimensional space. For this purpose we used extensively investigated family of recurrent artificial neural networks (RNN) called “Echo state network” (ESN). ESN incorporates a randomly generated recurrent reservoir with sigmoid nonlinearities of neurons outputs. The procedure called Intrinsic Plasticity (IP) that is aimed at reservoir output entropy maximization was applied for adapting of reservoir steady states to the multidimensional input data. Next we consider all possible combinations between steady states of each two neurons in the reservoir as two-dimensional projections of the original multidimensional data. These low dimensional projections were subjected to subtractive clustering in order to determine number and position of data clusters. Two approaches to choose a proper projection among the all possible combinations between neurons were investigated. The first one is based on the calculation of two-dimensional density distributions of each projection, determination of number of their local maxima and choice of the projections with biggest number of these maxima. The second one applies clustering to all projections and chooses those with maximum number of clusters. Multispectral data from Landsat 7 Enhanced Thematic Mapper Plus (ETM+) instrument are used in this work. The obtained number and position of clusters of a multi-spectral image of a mountain region in Bulgaria is compared with the regional landscape classification.

Dynamic Sound Fields Clusterization Using Neuro-Fuzzy Approach

In the presented investigation a recently proposed approach for multidimensional data clustering was applied to create a 3D “sound picture” of the data collected by a microphone array antenna. For this purpose records of acoustic pressure at each point (a microphone in the array) collected for a given period of time were used. Features for classification are extracted using overlapping receptive fields based on the model of direction selective cells in the middle temporal (MT) cortex. Next the clustering procedure using Echo state network and subtractive clustering algorithm is applied to separate these receptive fields into proper number of classes. Obtained for each time step two dimensional “sound pictures” were combined to create a 3D representation of dynamic changes in the sound pressure. We compare our results with the sonograms created by the original software of the producer of microphone array. Although our approach did not account for the distance to the noise source, it allows consideration of dynamically changing sounds.

An Accelerated IMM JPDA Algorithm for Tracking Multiple Manoeuvring Targets in Clutter

Theoretically the most powerful approach for tracking multiple targets is known to be Multiple Hypothesis Tracking (MHT) method. The MHT method, however, leads to combinatorial explosion and computational overload. By using an algorithm for finding the K-best assignments, MHT approach can be considerably optimized in terms of computational load. A much simpler alternative of MHT approach can be the Joint Probabilistic Data Association (JPDA) algorithm combined with Interacting Multiple Models (IMM) approach. Even though it is much simpler, this approach can overwhelm computations as well. To overcome this drawback an algorithm due to Murty and optimized by Miller, Stone and Cox is embedded in IMM-JPDA algorithm for determining a ranked set of K-best hypotheses instead of all feasible hypotheses. The presented algorithm assures continuous maneuver detection and adequate estimation of manoeuvring targets in heavy clutter. This affects in a good target tracking performance with limited computational and memory requirements. The corresponding numerical results are presented.

Improving super-resolution image reconstruction by in-plane camera rotation

In a digital optical imaging system, image resolution is constrained by several factors, including focus plane array pitch and optics. Super-resolution approaches aim at overcoming some of these limits by incorporating additional information of the object and/or combining several pictures of the same object, taken with some displacements between each other. This paper considers the second class of methods. The obtainable resolution improvement in this case has an upper limit, determined by the signal-to-noise ratio of the image taken. Moreover, some Fourier spectrum frequencies below this limit are unrestorable. Here an approach is introduced to overcome this by active control of the camera movements. An experiment, verifying the approach is presented.

Smart Feature Extraction from Acoustic Camera Multi-sensor Measurements

The paper applies recently developed smart approach for feature extraction from multi-dimensional data sets using Echo state networks (ESN) to the focalized spectra obtained from the acoustic camera multi-sensor measurements. The aim of the study is development of distance diagnostic system for prediction of wearing out of bearings. The procedure for initial features selection and features extraction from the focalized spectra was developed. Then the k-means clustering algorithm and Support vector machine (SVM) classifiers were applied to differentiate the tested bearings into two classes with respect to their condition (“Good” or “Bad”). The results using different dimensions of the extracted features space were compared.

ACD with ESN for Tuning of MEMS Kalman Filter

In the present work we designed a neuro-fuzzy approach for on-line optimal tuning of a Kalman filter of a gyroscope within a Micro ElectroMechanical Sensor (MEMS) device. It consists of Adaptive Critic Design (ACD) scheme in which the controller (a Fuzzy Rule Base (FRB) designed to adapt the measurement noise covariance matrix of a Kalman filter) is tuned using only information about the direction to which the estimation error changes (increase or decrease). A novel fast training dynamic neural network structure - Echo state network (ESN) - was used in the role of the critic element. Application to data collected from real MEMS demonstrated the ability of the proposed approach to tune Kalman filter and improve the quality of its estimates in changing working conditions of the MEMS in real time.

Multiple Hypothesis Tracking Using Hough Transform Track Detector

The Multiple Hypothesis Tracking algorithm (MHT) is an effective algorithm for moving objects detection and tracking.1,2 Few versions of this complex algorithm are described and evaluated in 1,2,4. Its measurement oriented version is considered as the most effective from theoretical point of view, but its practical implementation is limited because of the required significant computational load in cluttered environment. Several techniques minimizing this load were proposed,1,2,4 but they do not provide general solution to these problems. A new problem solution is proposed in this paper. A Hough Transform (HT) track detector is used for preliminary filtering of arriving false alarms (FA). The tracks detected in this way are processed asynchronously with another standard MHT algorithm to include them in the overall MHT scheme. The standard and the proposed MHT-HT algorithm (MHT2-HT) are evaluated and compared in the paper. The proposed algorithm shows remarkably good performance in cluttered environment at the cost of delayed track detection process.